Automated Interstudy Image Registration Techniquefor SPECT and PETStefan Eberl, Iwao Kanno, Roger R. Fulton, Anneke Ryan, Brian F. Hutton and Michael J. FulhamDivision of Medical Imaging Services, Scientific and Engineering Services Group; PET Department, Royal Prince AlfredHospital; and School of Physics, University of New South Wales, Sydney, Australia; and Department of Radiology andNuclear Medicine, Research Institute for Brain and Blood Vessels, Akita City, Japan

We report the extended application of an automated computertechnique for three-dimensional spatial registration of SPECT andPET studies. Methods: The technique iteratively reslices a misaligned data set until the sum of the absolute differences (SAD) froma reference data set is minimized. The registration accuracy wasassessed in Hoffman brain phantom studies collected with knownmisalignments and transmission studies of a thorax phantom withfiducial markers. The SAD was compared with three other costfunctions: stochastic sign change criterion, sum of products andstandard deviation (s.d.) of ratios. In clinical neurological and myo-cardial perfusion studies, registration accuracy was estimated fromthe relative locations of landmarks in the reference and registereddata sets. Results: Registration accuracy in the Hoffman brainphantom studies was -0.07 ±0.46 mm (mean ±s.d.) for translations and -0.01 ±0.20°for rotations, with maximum translation and

rotation errors of 1.2 mm and 0.8 degree, respectively. The SAD wasthe most accurate and reliable cost function. Registration errors inthe thorax phantom were 3.1 ±1.7 mm. Mean accuracy in theneurological studies, estimated from landmark pairs, was 2.0 ±1.1mm for SPECT to SPECT and 1.8 ±1.1 mm for PET to SPECTregistrations. Average registration accuracy in 201TImyocardial per

fusion studies was 2.1 ±1.2 mm. Conclusion: Our registrationmethod (a) provided accurate registrations for phantom and clinicalSPECT and PET studies, (b) is fully automated, (c) simplifies comparison of data sets obtained at different times and with differentmodalities, and (d) can be applied retrospectively.

Key Words: PET; SPECT; intermodalityregistration;image comparisonJ NucíMed 1996; 37:137-145

In functional imaging there is often a need to comparetomographic patient studies acquired at different times. Aprerequisite for accurate comparison is precise three-dimensional registration of data sets. Relatively small, but nonethelesssignificant, misalignments may occur in sequential examinations, such as PET activation studies using l5O-water or

dynamic PET acquisitions. Large misalignments can occur inexercise-redistribution 20lTl SPECT studies, WmTc-hexam-

ethyl-propyleneamine oxime (HMPAO) ictal/interictal brainperfusion studies in patients with epilepsy, other longitudinalPET or SPECT studies and in intermodality (SPECT to PET)comparisons (1-5).

Manual reslicing of data using functional-anatomic landmarks is frequently used to register reconstructed studies,particularly in SPECT. Disadvantages of such methods areoperator subjectivity and the time required to correct for the 6°

of motion (three linear and three angular displacements).Semiautomated techniques lessen this subjectivity, but theymay still be time-consuming or may not correct for all 6°of

freedom (4). Techniques that match surface contours have

Received Oct. 31, 1994; revision accepted May 15, 1995.For correspondence or reprints contact: Stefan Eberi, MSc, Department of Nuclear

Medicine, Royal Prince Alfred Hospital, Missenden Road, Camperdown 2050, Australia.

found applications in registering MRI and CT studies to PETexaminations but often require manual editing of data togenerate appropriate surfaces and have so far been largelylimited to brain studies (6-11). Typically, differences in activ

ity distributions in sequential SPECT and PET studies using thesame radiotracer are relatively small, and thus the functional-anatomic information contained in the voxel data can be usedfor registration. Techniques that use voxel data may also beapplicable across modalities (e.g., SPECT to PET) or fordifferent radiopharmaceuticals if the distributions of the respective radiotracers are sufficiently similar. A number of fullyautomated techniques based on voxel data have been described(1,3,12,13). One of these has also been applied to non-neurological PET studies (13). Except for the technique described by Woods et al. (14) for MRI to PET registration, thesetechniques have only been validated for registration errors inPET.

In this article, we describe and validate a fully automatedmethodology that registers SPECT to SPECT, PET to PET andSPECT to PET studies of the heart, brain and other organsystems that is based on functional-anatomic information

contained in voxel data. In a preliminary article (3), we reportedthe application of this technique to correct for PET interstudymotion in 15O-water regional cerebral blood flow activation

studies.

MATERIALS AND METHODS

Image Registration AlgorithmThe registration technique is based on the assumption that a

"cost function" can be defined, based on the voxel data, which

reaches an optimum when the two studies are perfectly registered.The sum of the absolute pixel-by-pixel differences (SAD) approaches a minimum for registered studies and thus fulfils therequirements for the cost function (3,4,13). The algorithm isdepicted in Figure 1. One study is arbitrarily designated thereference study, and the second study is registered to it. Themisaligned study is resliced with initial estimates of x, y and ztranslations and transaxial, coronal and sagittal rotations. Theresliced data are compared with the reference study, and the SADvalue is calculated. On the basis of this value, a new set of resliceparameters are estimated by the Simplex multidimensional minimization routine (/5). The resliced data are again compared withthe reference data, and the cost function is recalculated. Thisprocedure is repeated until the sum of the absolute differences isminimized.

The following published cost functions were also implementedand tested: (a) stochastic sign change (SSC) criterion, which countsthe number of sign changes in the subtraction data (13,16-18); (b)sum of pixel-by-pixel products (SProd) (19); and (c) standarddeviation of pixel-by-pixel ratios (SDralio)(1,14).

Global activity differences between the studies are normalized ateach iteration using the global means of the two studies before the

IMAGEREGISTRATIONFORSPECT ANDPET •Eberl et al. 137

by on July 25, 2020. For personal use only. jnm.snmjournals.org Downloaded from

FIGURE 1. Schematic representation ofregistration algorithm. At each iteration,misaligned data are resliced and compared with reference data, and the SAD(Ildiffl) cost function value is calculated.The Simplex routine provides new reslicing coefficients at each iteration until aminimum in the cost function is reached.

Misaligned Study Reference Study

Initial values

for Simplex

//

Resliced Data

Õ

Subtract fromReference Study

New ReslicingT3 Translations &Coefficients I 3 Rotations I Difference

Images

CalculateIldiffl

Cost Function

UpdateReslicing

Coefficients

MinimumCost Function

9

CostFunctionValue

AUGNMENT COMPLETED

cost function is evaluated. Only voxels with counts above auser-specified threshold contribute to the cost function. The threshold value is not critical and has been determined empirically foreach type of study as a fraction of the maximum voxel count in thereference study (0.2 for brain, 0.15 for heart and 0.0 for regionalwhole-body studies). The speed of registration is improved bycomputing the cost function from every fourth or eighth voxel,depending on the number of slices in the study. Trilinear interpolation is used for reslicing. Interpolation to isotropie voxel size isperformed before the registration procedure. Differences in matrix

dimensions, number of slices and voxel sizes between the referenceand misaligned studies are automatically adjusted to allow registration of data obtained from different imaging systems or withdifferent acquisition and reconstruction parameters.

Experimental DesignStudies conducted to evaluate the technique are summarized in

Table 1. The technical accuracy of the technique was assessed inPET studies of a three-dimensional Hoffman brain phantom (20)that were collected with known misalignments. These studies were

TABLE 1Summary of Evaluation Studies

Study type Acquisition/Registration Registration evaluation

Phantom studiesHoffman brain phantom

Thorax phantomPatient Studies

Brain SPECTBrain SPECT/PET

Cardiac SPECT

18F-PET acquisition with known misregistrations and

artificial lesionsRegistration of SPECT-PET transmission studies

Registration of pairs of ""Tc-HMPAO SPECT studiesRegistration of 18F-FDG PET to "Tc-HMPAO SPECT

studiesRegistration of ^"Tl myocardial perfusion SPECT studies

Comparison of registration parameters to knownmisregistration

Differences in fiducial marker locations

Differences in landmark locationsDifferences in landmark locations

Differences in landmark locations

138 THE JOURNALOFNUCLEARMEDICINE•Vol. 37 •No. 1 •January 1996

by on July 25, 2020. For personal use only. jnm.snmjournals.org Downloaded from

TABLE 2SPECT Acquisition and Reconstruction Parameters

Studycamera201Tl-heart

(Triad XLT)Ceretec brain

Pad)Triad XLTCollimatorHigh-resolution

parallelHigh-resolution

parallelUltra-high-

resolution fanbeamTotal

no. ofprojections90°

120°

120°Time

perprojection

(sec)30

30

40Acquisition

matrix128x64

128 x 128

256 x 128PrefilterBW

fc = 0.40,n = 10BW fc = 0.60,n = 8BW fc = 0.60,n = 8Reconstruction

matrix64

x64

64X64

128 x 128Pixel

size(mm)3.6

5.3

1.8

Total number of projections are given. Number of projections collected per detector is total number divided by number of detectors (three for Triad, twofor Biad).

BW = Butterworth filter; fc = cutoff frequency (cycles/cm); n = order of the Butterworth filter.

also used to compare the different cost functions; SPECT and PETtransmission studies of a thorax phantom with fiducial markerspermitted assessment of accuracy in non-neurological studies.

Registration accuracy in clinical neurological and myocardialSPECT and PET studies was estimated from distances betweeneasily identifiable landmark pairs. The patients selected for thisassessment were chosen randomly from routine clinical studies.

InstrumentationAll PET studies were performed on an ECAT-951R whole-body

tomograph (Siemens/CTI, Knoxville, TN), a 16-ring system, whichsimultaneously acquires 31 slices with a slice separation of 3.38mm. Studies were reconstructed in a 128 X 128 matrix. SPECTtransmission studies were carried out on a Philips Diagnosi Agamma camera (Philips Medical Systems, Hamburg, Germany)with a scanning transmission line source (21). Clinical SPECTstudies were collected on a Trionix TRIAD XLT triple-head orBIAD-24 dual-head gamma camera (Trionix, Twinsburg, OH). TheSAD cost function was used for registering all clinical studies.Collection and reconstruction parameters for the clinical SPECTstudies are summarized in Table 2. Sun Spare work stations (SunMicrosystems, Mountain View, CA) were used for all registrations.The software was written in C reading directly either ECAT PETor Trionix SPECT file format. Interfile was used as an intermediateformat to convert files for PET to SPECT registrations (22).

Phantom StudiesHoffman Brain Phantom. The Hoffman brain phantom (20) was

filled with 340 MBq (0.26 MBq/ml) of I8F and scanned at variousknown misalignments from a reference study on the ECAT-951Rscanner. Total data acquisition time was 15 min. Data werereconstructed with a Hann reconstruction filter (cutoff frequencyfc = 0.5 cycles/pixel) and zoomed to give pixel sizes of 2.4 or 1.9mm. Calculated attenuation correction with automatic contourdetection was used (23). Misalignments were produced throughgantry tilt, rotation and vertical and horizontal bed motions.Rotations of up to 16°and translations up to 15 mm were

introduced. Translations and rotations determined by the programto achieve registration were compared with the known misalignments for each cost function. Sixteen phantom studies wereregistered and analyzed in this manner.

Changes in activity distribution between studies were simulatedby implanting single spherical lesions (radii up to 2 cm), withactivity concentrations increased by up to 45% of the originalactivity concentration, into grey matter regions in one of thephantom studies before registration. A total of 18 lesion studieswere generated, and registration accuracy was compared with thatobtained without lesions.

Thorax Transmission Phantom. PET and SPECT transmissionstudies of a thorax phantom were performed. The PET transmissionstudies were collected for 20 min on the ECAT-951R using threerotating 68Ge/68Garod sources. The SPECT data were collected for

40 sec at each of the 64 projection angles using 64 x 64 matriceswith a scanning transmission line source (21 ) containing approximately 4 GBq of 99rnTc.Reconstructed full width at half maximum

resolution was 10 mm for the PET study and 17 mm (in-plane) forthe SPECT transmission study. The SPECT transmission study wasregistered with the three PET transmission studies of the samephantom collected at different positions. Each of the three PETtransmission studies was deliberately mispositioned. The studieswere registered with the SAD cost function. Final pixel size andslice thickness after registration were 3.1 mm for both the PET andSPECT studies. Small rectangular regions of interest (ROIs) wereplaced over each fiducial marker in the transaxial slices. For eachROI, the pixel-weighted count centroid was calculated over allpixels in the ROI to give the fiducial marker coordinate. Registration accuracy was estimated from the distances between fiducialmarker coordinate pairs as follows: If (x,, y,) are the coordinates ofthe fiducial marker centroid on the reference slice and (x2, y2) thecentroid coordinates on the corresponding registered slice, then

Distance = + Ay2, Eq. 1

where Ax = x2 - x,, and Ay = y2 - y,.

To assess the reproducibility in locating centroids, studies wereanalyzed on two separate occasions, several weeks apart. Thedistance between corresponding fiducials from the first and secondanalyses were calculated using Equation 1 for both the SPECT andPET studies. In this case, (x,, y,) were the coordinates of themarker in the first analysis and (x,, y2) the coordinates of the samemarker in the same study from the second analysis.

Patient StudiesBrain SPECT. Nine pairs of brain SPECT examinations were

used for the assessment. Eight of these patients were referred forassessment of refractory epilepsy and thus had ictal and interictalstudies. The remaining patient had cerebral lupus. The SPECTacquisitions were carried out at least 15 min after injection of 550MBq of 99mTc-HMPAO. All data were reconstructed with calculated attenuation correction (/x = 0.11 cm"'). The Triad XLT

fan-beam data were rebinned into parallel projections beforeprefiltering, reconstruction and attenuation correction. Both interictal HMPAO and interictal fluorodeoxyglucose (FDG) PET studies were registered with the ictal HMPAO study.

Brain FDG-PET. The FDG-PET studies were available in sevenof the nine brain SPECT examinations used for the assessment. The

IMAGEREGISTRATIONFORSPECT ANDPET •Eberl et al. 139

by on July 25, 2020. For personal use only. jnm.snmjournals.org Downloaded from

TABLE 3Three-Dimensional Hoffman Brain Phantom Studies Registration

Errors

Translation errors (mm) Rotation errors (degree)

CostfunctionSADSSCSProdscu,Mean-0.07-0.72(-0.16-0.47(-0.36-0.95(-0.04-t--t-±•+••+•-t-±-+-s.d.(n=0.461.570.31)4.070.32)2.360.31)Maximum48)1.15-7.08(-0.84)-11.94(-0.62)-8.06(0.96)Mean-0.011.78(-0.140.75(-0.602.31(0.01-I--H-t--1--*--+-±-f-s.d.(n=0.205.300.48)4.780.66)5.560.30)Maximum48)0.8125.49(-1.21)18.09(-1.36)23.37(1.24)

Values in parentheses exclude the unsuccessful registrations.n is the number of values averaged, that is, three translations (x, y and z)

or three rotations (transverse, coronal and sagittal) from each of 16 registrations.

FDG-PET studies were collected for 30 min after a 30-min uptakeperiod after injection of 370 MBq [I8F]FDG. The data werereconstructed with a Shepp filter (fc = 0.5 cycles/pixel) and a zoomof 2.55 to give a pixel size of 1.8 mm. Measured attenuationcorrection was performed. PET studies were registered to the ictalSPECT studies.

Cardiac SPECT. Myocardial 2U1T1perfusion studies were per

formed on the Triad XLT gamma camera according to theparameters in Table 2. Data from a 180°arc (left posterior oblique

to right anterior oblique) were reconstructed. Redistribution (eightpatients) and reinjection (five patients) scans were registered to theexercise or 0 hr studies. All studies were performed according toour standard clinical exercise or resting myocardial perfusionprotocols.

Assessment of Registration Accuracy in Patients. A techniquesimilar to that described for the transmission phantom studies wasused to estimate registration accuracy in patients. Small rectangularROIs were centred on functional-anatomic "landmarks," instead

of external fiducials. Distances between landmark pairs in clinicalstudies were again calculated with Equation 1, similar to thetransmission phantom studies. Five landmarks were comparedindependently for each of the three orthogonal slices for a total of15 landmarks per study set. The landmarks were spread over thewhole image volume, not just over the central region. Distinctlandmarks were less easily identified in the myocardial studies.Thus, ROIs were placed on the apex, sections of the myocardialwall, near the right ventricular attachment point or at the lateral,septal, inferior, anterior or basal limits of the myocardium. Displayed slices were magnified (pixel sizes were 1.5 mm for brainand 1.8 mm for cardiac studies) to assist in accurate placement ofthe ROIs.

RESULTS

Phantom StudiesHoffman Brain Phantom Studies. Technical accuracy for the

Hoffman brain phantom studies is summarized in Table 3. TheSAD cost function achieved the best results and registered all16 studies satisfactorily with overall accuracy of 1.5 mm and0.8°.The SSC cost function satisfactorily registered 5 studies,

the SProd cost function 1 and the SDraljo cost function 10 of the16 studies. Satisfactorily registered studies had registrationerrors similar to those observed with the SAD cost function(Table 3). Unsatisfactory registrations were usually visually

FIGURE 2, Representative slices from registrationof Hoffman brain phantom PET studies. Top row, reference study. Second row, translational androtational misregistrations are apparent in the misaligned study. Third row,slices after registration. In bottom row, subtraction images (registered study-

reference study) show no significant misregistration artifacts.

obvious and had large errors. Figure 2 illustrates a typicalexample of accurate registration with the SAD cost function.Equally spaced slices through the upper, middle and lowersections of the phantom (slices 18,29 and 40, respectively, froma total of 55 slices in the phantom study) are displayed. Noappreciable misalignment artifacts are apparent in the subtraction images after registration.

No systematic registration errors were introduced by theimplanted lesions. Mean differences in calculated translationand rotation coefficients for studies with lesions compared withthe same study without a lesion were 0.01 ±0.02 mm (mean ±s.d.) and 0.00 ±0.06°,respectively.

Thorax Transmission Phantom. The mean distance betweenfiducial marker pairs after registration was 3.1 ±1.7 mm (n =

15, i.e., five fiducial markers in each of the three misregisteredstudies). The maximum distance was 5.9 mm. Figure 3 showsthe close agreement between the fiducial marker coordinates inthe SPECT and PET studies. Slices containing fiducial markers(slices 10 and 15) and the inferior extent of the "lung" (slice

20) confirm the registration accuracy (Fig. 4). The followingresults were obtained for the reproducibility in locating themarkers. The mean distance between the first and secondanalysis was 1.4 ±0.9 mm (n = 15) for the PET studies and 2.6±1.3 mm (n = 15) for the SPECT reconstructions. For the

combined SPECT and PET studies, the mean distance was 2.0± 1.3 mm (n = 30). Maximum distances of 3.2 mm were

observed for the PET and 4.5 mm for the SPECT and combinedstudies.

140 THEJOURNALOFNUCLEARMEDICINE•Vol. 37 •No. 1 •January 1996

by on July 25, 2020. For personal use only. jnm.snmjournals.org Downloaded from

M

4035

ft 20

»-•0.5554 t 1.0068 x

R : 0.999«

20 40 «O 80PET Study Fiducial Co-ordinate

FIGURE 3. Plot of thorax transmission phantom results. The fiducial markercentroid x and y coordinates of registered SPECT slices are plotted againstthe coordinates of the corresponding markers in the PET slices. Coordinatesare in pixels from the 128 x 128 transverse slices.

Patient ResultsBrain SPECT and PET Studies. Distances between landmark

pairs after registration are plotted in Figure 5. Overall meandistances between landmark pairs for transverse, coronal andsagittal slices were 2.0 ± 1.1 mm (n = 135) for SPECT toSPECT registrations and 1.8 ±1.1 mm (n = 105) for PET toSPECT registrations. Maximum distances of 4.8 mm wereobserved for SPECT to SPECT registrations and 5.1 mm forPET to SPECT registrations. Satisfactory registration of ictal,interictal SPECT and interictal PET studies was achieveddespite the marked focal increase of tracer uptake in the ictalSPECT study. This registration is illustrated in Figure 6, whichcontains slices through the epileptic focus (middle column) aswell as superior and inferior slices to cover most of the brain.

Cardiac SPECT Studies. Figure 7 shows a plot of thedistances between landmark pairs after registration in eightpatients with myocardial perfusion studies. Five of the eight

patients had reinjections in addition to the redistribution studies.Registration results for both redistribution and reinjectionstudies are combined in Figure 7. The overall mean distancewas 2.1 ±1.2mm(n= 195), with a maximum error of 5.8 mm.

Other Clinical Studies. The technique is successfully applied ona routine basis to other clinical SPECT and PET studies, such as67Ga-citrate and regional whole-body PET studies. Satisfactory

registration of FDG PET regional whole-body acquisitions to67Ga-citrate studies is possible with the technique despite differ

ences in activity distribution between the two tracers (Fig. 8).Slices through the anterior extent of the liver and the tumor sitesclearly demonstrate the level of registration achieved.

The time taken for the registrations depended on the matrixsize, number of slices in the studies and number of iterationsrequired to achieve registration. Registration times, using a SunSpare 10/41 work station, ranged from 2 min for a 20IT1 study

(64 X 64 matrix) to 15 min for a brain study (128 X 128 matrix).For total-body studies, with nearly 200 slices, the registration timeis typically 45 min. Since the technique is fully automated, it canbe left running unattended without operator intervention, exceptfor the initial setup which requires 1-2 min.

DISCUSSIONThe main findings from our study are:

1. Our registration technique achieved a high degree ofaccuracy.

2. The SAD cost function was the most accurate costfunction in the brain phantom studies.

3. The technique is applicable to functional imaging studieswith similar tracer distributions between studies.

4. The technique is relatively unaffected by localized differences in activity distribution and the range of noise inclinical studies, such as interictal and ictal HMPAO-SPECT and FDG-PET brain studies and exercise/redistribution 201T1myocardial perfusion studies.

Registration and Cost Function AccuracyIn Hoffman brain phantom studies, our technique attained a high

degree of registration accuracy. Localized differences in activity

FIGURE 4. Representative slices ofthorax phantom. Top row shows slicesfrom a reference PET transmissionstudy with registered SPECT study below. Fiducial markers are indicated bywhite arrows. Satisfactory registration isseen, despite differences in resolutionand noise between PET and SPECTstudies.

IMAGEREGISTRATIONFORSPECT ANDPET •Eberl et al. 141

by on July 25, 2020. For personal use only. jnm.snmjournals.org Downloaded from

o§3= 3Ö

l 2

FIGURE 5. Distances, in millimeters,between landmark pairs after registration of""Tc-HMPAO SPECT to SPECT studies(solid diamonds) and FDG-PET to "Tc-

HMPAO SPECT data (open diamonds).Data are shown for landmarks in thetransverse, coronal and sagittal slices.Mean ± s.d. is also plotted for eachorthogonal slice set (squares and verticallines).

o

-»

Transverse CoronalSlices

Sagina!

FIGURE 6. Registered SPECT and PETstudies in a patient with intractable focalepilepsy. Reference ictal SPECT study(top row) shows increased perfusion inthe left temporal lobe, with correspondingslight reduction in registered interictalSPECT study (middle row). Registeredinterictal FDG-PET scan (bottom row)shows left temporal lobe glucose hypo-

metabolism.

142 THEJOURNALOFNUCLEARMEDICINE•Vol. 37 •No. 1 •January 1996

by on July 25, 2020. For personal use only. jnm.snmjournals.org Downloaded from

65E,-_

miI£

3^"0C.2

,73810|-r

........ , , ,v-

*»*; :• *t.*,*•«::t: ^^ ** I !*i*»^

T**: *> **^ <#** *f<H4*^\

»^ " *^*I; C" 4̂¿** ^x*Short

Vertical Long HorizontalLAxisSlices----•-.i-ong

FIGURE 7. Distances between landmarkpairs in 201T1myocardial perfusion studies

after registration. Data for both redistribution and reinjection studies registered toexercise (or 0 hr) studies are included inthe plot. Mean ±s.d. is also plotted foreach of the orthogonal slice sets.

distributions and moderate levels of noise had a minimal effect onthe final accuracy. The accuracy of SPECT to PET transmissionphantom registration shows that the technique is not limited toneurological emission studies. The larger errors observed in thetransmission phantom experiment are at least partly attributable toerrors in localizing the fiducial markers. In fact, the registrationerrors are only marginally larger than the reproducibility inlocating the markers in the SPECT studies. Thus, the registrationerrors are considered to be conservative estimates.

We found that the SAD cost function was the most reliable costfunction. The other cost functions had varying degrees of successin accurately registering the phantom studies. In some studies,registration accuracy approached that of the SAD, but with others

there were large errors and clear misregistrations. These resultssuggest that the optimization routine was trapped by a localoptimum (maximum or minimum depending on cost function). Itwas thus prevented from reaching the global optimum of properlyregistered studies. Hoh et al. (13) avoided this problem with theSSC cost function by using the final reslicing parameters of theSAD registration as starting values for the SSC registration. Thus,the SSC registration started near its global optimum, which reducesthe likelihood of convergence to a local optimum, which may beencountered when optimizing the SSC from large initial misregistrations. Despite the theoretical advantage of the SSC, Hoh et al.(13) concluded that the SSC did not significantly improve theaccuracy achieved with the SAD. We had less success with the SDrjtjo

FIGURE 8. Coronal slices of 67Ga-citrate SPECT study of a patient with gastric lymphoma (top row) with slices of the registered regional whole-body FDG-

PET study (bottom row). Despite marked regional differences in activity distribution, satisfactory matching of the data sets is apparent. Direct comparison of67Ga-citrate and FDG uptake in multiple tumor sites (arrows) is possible. (Note: Arms were excluded in SPECT acquisition but present in PET field of view.)

IMAGEREGISTRATIONFORSPECT ANDPET •Eberl et al. 143

by on July 25, 2020. For personal use only. jnm.snmjournals.org Downloaded from

cost function than that reported by Woods et al. (1,14), who used adifferent minimization routine. There may have been differences in thedetailed implementation of the cost function to explain our lowersuccess rate with their cost function. Our results do suggest, however,that the SAD cost function is relatively free of local minima comparedwith the other cost functions.

For clinical studies, we estimated the registration accuracydirectly from the image data by localizing corresponding landmarks inthe reference and registered studies. This step avoids problems withfiducial markers, which are rarely reliable if the studies are weeks ormonths apart and which do not always have a fixed relationshipbetween the organ of interest and the external markers in, for example,cardiac studies. As we reported, however, for the transmission phantom analysis, there is an uncertainty associated in locating thelandmarks. Thus, we consider that our registration error estimates areconservative. Despite this, we observed an average landmark registration error of about 2 mm for both neurological and cardiac clinicalstudies.

LimitationsVoxel-based registration techniques assume that the activity

distribution of the studies being registered is similar. Our data andthat of others (1,13,18) show that even large focal differences inactivity distributions have little effect on the final registrationaccuracy. Thus, registration is not prevented by tracer distributiondifferences in clinical studies acquired with the same radiophar-

maceutical but at different times (e.g., interictal/ictal HMPAObrain studies and exercise/redistribution 20IT1 myocardial perfusion

studies).When different radiotracers were used, the performance of our

technique depended on the relative contributions that regions withdisparate tracer uptake made to the cost function. For example,relatively higher FDG uptake, compared with 67Ga uptake, was seen

in the heart and right kidney (Fig. 8). The cost function value fromthese organs could be reduced by incorrectly reslicing the PET data tooverlap the heart or kidney from the FDG-PET study with the liver ofthe 67Ga SPECT study. This reslicing would, however, increase the

cost function because of misregistration of areas with congruentuptake (liver, rumor, etc). Registration was successful because areaswith corresponding uptake (e.g., liver, tumor) dominated the costfunction over incongruent uptake regions (heart, kidney). The successof our technique in registering data sets with different radiotracersdepends on physical size, relative uptake and location relative to otherstructures of regions with disparate uptake. There is no definite cutoffpoint; rather, the more marked the tracer distribution differences themore increased the likelihood of trapping in local minima andmisregistration. For certain combinations of radiotracers, however, ourtechnique consistently misregistered the data. For example, in patientswith brain tumors, our technique registered the high scalp uptake ofthe 20IT1 brain SPECT study mainly to the cortical uptake of FDG in

superior regions of the brain. In addition, our technique is also unlikelyto correctly register "mTc bone SPECT studies to 67Ga SPECT or

FDG-PET studies. For these study combinations, transmission studies,

if performed, could be used for registration of the emission data.

Clinical ApplicationsWe apply our technique routinely to intramodality registration of

sequential HMPAO brain, 201T1 myocardial perfusion and 67Ga

SPECT studies, as well as FDG brain, regional whole-body and'5O-water PET studies. Across modalities, we use the technique forregistering HMPAO-SPECT and FDG-PET brain studies, 201T1

SPECT and FDG/NH3 PET myocardial viability studies, as well as67Ga SPECT and FDG-PET regional whole-body studies. The

technique produces accurate registration for direct visual comparison and eliminates operator time required to manually reslice andmatch slices for review. Parametric images (e.g., subtraction

images) derived from registered studies, however, should beinterpreted with caution. Even with perfect registration, partialvolume effects, attenuation correction inaccuracies and differencesin resolution and count statistics may cause artifacts in parametricimages, particularly in SPECT studies. Registration techniquesshould not be regarded as a substitute for care and consistency inpatient positioning. Attention to consistent and reproducible patientpositioning minimizes registration errors while providing optimumresults from registration of multiple patient studies.

CONCLUSIONOur automated image registration technique is accurate in

both phantom and clinical studies. It can be successfully appliedto a range of SPECT and PET studies as well as PET to SPECTregistrations. Its routine clinical use in our department greatlyfacilitates direct comparison of longitudinal SPECT and PETstudies. It requires minimal operator time and virtually eliminates the need for reanalysis of data. The technique successfullyregisters studies with marked local differences in activitydistributions. Transmission data may be used where directregistration based on emission data is not possible because ofdistribution differences of the radiopharmaceuticals used.

ACKNOWLEDGMENTSWe thank Steven Meikle and Gerhardt Lowenthal for helpful

comments and discussions. This study was supported in part bythe Targeted Institutional Links Project grant between University of Technology, Sydney, Royal Prince Alfred Hospital andResearch Institute for Brain and Blood Vessels, Akita.

REFERENCES1. Woods RP, Cherry SR. Mazziotta JC. Rapid automated algorithm for aligning and

reslicing PET images. J Compiti Assist Tomogr 1992;16:620-633.

2. Links JM. The influence of positioning on accuracy and precision in emissiontomography [Editorial]. J NucíMed 1991:32:1252-1253.

3. Eberi S, Kanno 1. Fulton RR, et al. Automatic three-dimensional spatial alignment forcorrecting interstudy patient motion in serial PET studies. In: Proceedings of the BrainPET'93 Akita: Quantification of Brain Function. Akita, Japan. Amsterdam: Eiservier;

1993:419-426.4. Phillips RL, London ED. Links JM, Cascella NG. Program for PET image alignment:

effects on calculated differences in cerebral metabolic rates for glucose. J NucíMed1990:31:2052-2057.

5. Pate BD. Snow BJ, Hewitt KA, Morrison KS. Ruth TJ. Calne DB. The reproducibilityof striatal uptake data obtained with positron emission tomography and flourine-18-L-6-fluorodopa tracer in non-human primates. J NucíMed 199I;32:1246-1251.

6. Levin DN, Pelizzari CA, Chen GTY, Chen CT, Cooper MD. Retrospective geometriccorrelation of MR, CT and PET images. Radiology 1988:169:817-823.

7. Pelizzari CA, Chen GTY, Speibring DR, Wcichselbaum RR, Chen CT. Accuratethree-dimensional registration of CT, PET and/or MR images of the brain. J CompaiAssist Tomogr 1989;13:20-26.

8. Pietrzyk U, Herholz K, Heiss WD. Three-dimensional alignment of functional andmorphological tomograms. J Compiti Assist Tomogr 1990:14:51-59.

9. Holman BL. Zimmerman RE. Johnson KA, et al. Computer-assisted superimpositionof magnetic resonance and high-resolution technetium-99m-HMPAO and thallium-201 SPECT images of the brain. J NucíMed 1991:32:1478-1484.

10. Turkington TG, Jaszczak RJ, Pelizzari CA. et al. Accuracy of registration of PET,SPECT and MR images of a brain phantom. J NucíMed 1993:34:1587-1594.

11. Kapoules I, Alavi A, Alves WM, Our RE, Weiss DW. Registration of three-dimensional MR and PET images of the human brain without markers. Radiology1991:181:731-739.

12. Bacharach SL, Douglas MA. Carson RE. et al. Three-dimensional registration ofcardiac positron emission tomography attenuation scans. J NucíMed 1993:34:311-321.

13. Hoh CK. Dahlbom M, Harris G, et al. Automated iterative three-dimensionalregistration of positron emission tomography images. J NucíMed 1993:34:2009-2018.

14. Woods RP, Mazziotta JC, Cherry SR. MR1-PET registration with automated algorithm. J Compii! Assist Tomogr 1993:17:536-546.

15. Press WH, Flannery BP, Teukolsky SA, Vetterling WT. Numerical recipes in C: theart of scientific computing. New York: Cambridge University Press; 1988.

16. Venot A, Pranzato L, Walter E, Lebruchec JF. A distribution-free criterion for robustidentification, with applications in system modelling and image processing. Automatica 1986:22:105-109.

17. Venot A, Liehn JC, Lebruchec JF, Roucayrol JC. Automated comparison of scinti-graphic images. J NucíMed 1986:27:1337-1342.

18. Minoshima S, Berger KL, Lee KS, Mintun MA. An automated method for rotationalcorrection and centering of three-dimensional functional brain images. J NucíMed1992:33:1579-1585.

19. Junck L. Moen JG. Hutchins GD. Brown MB. KühlDE. Correlation methods for thecentering, rotation and alignment of functional brain images. J NucíMed 1990;3I:1220-1276.

144 THE JOURNALOFNUCLEARMEDICINE•Vol. 37 •No. 1 •January 1996

by on July 25, 2020. For personal use only. jnm.snmjournals.org Downloaded from

20. Hoffman EJ, Cutler PD, Digby WM, Maziotta JC. Three-dimensional phantom to 22. Cradduck TD, Bailey DL, Hutton BF, et al. A standard protocol for the exchange ofsimulate cerebral blood flow and metabolic images for PET. IEEE Tram Nucl Sci nuclear medicine image files. NucíMed Commun 1989;IO:703-713.1990;NS-37:6I6-620. ,.. D .... ..... , _ . - „, ,, „ r ,

-.1 T- r, r-, -i r*T «•11 em -m. i M «L-1 nn 11 r*r- « - i- 23. Bergstrom M, Litton J, Enksson L, Bohm C, Blomqvist G. Determination of object21. Tan P, Bailey DL, Meikle SR, Eberl S, Fulton RR, Mutton BF. A scanning line source

for simultaneous emission and transmission measurements in SPECT. J Nucl Med contour from projections for attenuation correction in cranial positron emission1993;34:1752-1760. tomography. J Comput Assist Tomogr 1982:6:365-372.

IMAGEREGISTRATIONFORSPECT ANDPET •Eberl et al. 145

by on July 25, 2020. For personal use only. jnm.snmjournals.org Downloaded from

1996;37:137-145.J Nucl Med.   Stefan Eberl, Iwao Kanno, Roger R. Fulton, Anneke Ryan, Brian F. Hutton and Michael J. Fulham  Automated Interstudy Image Registration Technique for SPECT and PET

http://jnm.snmjournals.org/content/37/1/137This article and updated information are available at:

  http://jnm.snmjournals.org/site/subscriptions/online.xhtml

Information about subscriptions to JNM can be found at:  

http://jnm.snmjournals.org/site/misc/permission.xhtmlInformation about reproducing figures, tables, or other portions of this article can be found online at:

(Print ISSN: 0161-5505, Online ISSN: 2159-662X)1850 Samuel Morse Drive, Reston, VA 20190.SNMMI | Society of Nuclear Medicine and Molecular Imaging

is published monthly.The Journal of Nuclear Medicine

© Copyright 1996 SNMMI; all rights reserved.

by on July 25, 2020. For personal use only. jnm.snmjournals.org Downloaded from